Today's cybersecurity news is dominated by one story: October 2026 was the worst month yet for MCP security. A CVSS 9.9 remote code execution in Langflow (CVE-2026-105697), an SSRF in fetch-server, a 9.3 in x64dbg, and credential theft in the Python and TypeScript MCP SDKs — and on October 18, OX Security reported scanning 15,465 public MCP servers, per October 2026 reporting. Keyword research lists "cybersecurity news today" at 550,000 monthly searches (Semrush organic keywords, securityweek.com overview): the world is watching, and the protocol that wires AI agents to tools is having its worst month on record.
October 2026 delivered a record cluster of Model Context Protocol vulnerabilities: Langflow RCE CVE-2026-105697 at CVSS 9.9, SSRF in fetch-server, a 9.3-severity flaw in x64dbg, and credential theft affecting the Python and TypeScript MCP SDKs, per October 2026 reporting. Days later, OX Security reported 15,465 public MCP servers on the internet — the blast radius for every one of those flaws.
Langflow RCE — CVE-2026-105697 (CVSS 9.9). Remote code execution on the AI workflow platform. A 9.9 is near-maximum severity: arbitrary commands on the host running your agent workflows.
fetch-server SSRF. Server-side request forgery in the fetch MCP server — an attacker can make the server request internal URLs, pivoting into private networks.
x64dbg (9.3). A high-severity flaw in the debugger tooling reachable through MCP integrations.
Python + TypeScript SDK credential theft. The official SDKs themselves — the code every MCP server is built on — had credential-theft issues, per October 2026 reporting.
Four hits in one month, across the platform layer, the tool layer, and the SDK foundation. That is why October earned the "worst month" label: it was not one bad bug, it was the whole stack at once.
Vulnerabilities are only half the story. OX Security's October 18 scan found 15,465 MCP servers reachable on the public internet. MCP was designed to connect agents to tools — but thousands of those connections are listening on network ports where anyone can probe them. Every unpatched server in that population is a potential entry point from the internet straight into someone's agent infrastructure. The fix starts with knowing whether you're one of them.
Scope note: this is a self-assessment questionnaire scoring your exposure to the October 2026 MCP CVE cluster — not a live vulnerability scan of your servers. Only patching and a real scan can confirm your status.
1. Do you run an MCP server?
2. Is it Langflow, fetch-server, x64dbg, or built on the Python/TypeScript SDKs?
3. When did you last update your MCP servers?
4. Which transport do your servers use?
5. Are secrets or credentials stored in the server config?
6. Do your agents connect to third-party or untrusted MCP servers?
Reporting: October 2026 industry coverage of the MCP CVE cluster; OX Security scan count (15,465 public servers, October 18). Technical details beyond the reported items not independently verified.
Rogue-Agent Defense — Ingenious AI's rogue-agent protection platform — watches for rogue agent behavior across your MCP surface, so a compromised server can't quietly turn your agents against you. Get early access on WhatsApp.
MCP — the Model Context Protocol — is the open standard that connects AI agents to tools, data sources, and services. An MCP server exposes capabilities (file access, API calls, databases) that an agent can invoke. Because servers bridge powerful models to real systems, a vulnerable MCP server is a direct path from an AI agent into your infrastructure.
The Langflow remote code execution flaw, CVE-2026-105697, rated CVSS 9.9 — about as severe as ratings get — per October 2026 reporting. Remote code execution on an AI workflow platform means an attacker can run arbitrary commands on the host, which in an agent infrastructure is close to a worst case.
On October 18, OX Security reported scanning 15,465 public MCP servers — meaning that many MCP servers were reachable on the public internet and observable. It is a measure of exposure at scale: the protocol that connects AI agents to tools is widely internet-facing, which multiplies the blast radius of every MCP vulnerability. Findings beyond the scan count were not part of the reporting reviewed for this page.
STDIO (local process) is the narrower attack surface: the server runs as a local subprocess with no network listener. SSE and HTTP transports expose the server on a network port, which is what makes the 15,465 publicly reachable servers possible. If your server does not need network access, STDIO local-only is the safer default.
Inventory every MCP server you run, patch the October cluster (Langflow, fetch-server, x64dbg, Python and TypeScript SDKs), prefer STDIO local transport, keep secrets out of server configs, and treat third-party MCP servers as untrusted input. Rogue-Agent Defense by Ingenious AI is built to watch for rogue agent behavior across this surface.